HTTP API

Every agentpack server exposes the same surface. With multiple packs loaded, address one via the pack query/body parameter (defaults to the first).

Endpoint	What it does
GET /	Dev UI — live network graph, token streaming, builder, themes
GET /api/health	{ ok, packs } — ok is true when an LLM key is configured
GET /api/packs	All loaded packs (including browser-built ones)
GET /api/network	{ nodes, edges } team topology
POST /api/run	Start a run: { query, pack?, threadId? } → { runId, threadId }
GET /api/stream/:runId	SSE stream of run events (/events/:runId alias kept for compatibility)
GET /api/runlog/:runId?after=N	Polling fallback — same events as plain JSON
POST /api/approve	Resolve a human-in-the-loop gate: { runId, approvalId, approve }
GET /api/tools	Tool registry with JSON schemas
POST /api/tools/execute	Direct tool execution — zero LLM tokens, deterministic
GET /api/toolcatalog	Every tool available to the builder
POST /api/packs	Create a custom team at runtime (no restart)
POST /api/suggest	AI-assisted builder: { description } → full team spec
GET /widget.js	Embeddable live network panel
ALL /mcp, /mcp/:pack	Auto-generated MCP server per team

Run a query end to end

RUN=$(curl -s -X POST http://localhost:3000/api/run \
  -H 'Content-Type: application/json' \
  -d '{"query": "Find healthcare software targets and score the best one."}' \
  | python3 -c "import json,sys; print(json.load(sys.stdin)['runId'])")

curl -N http://localhost:3000/api/stream/$RUN

Event kinds on the stream: run_started, hop, thinking, tool_executing, tool_complete, agent_step, answer_token, answer_chunk, usage, approval_request, approval_resolved, run_finished.

Conversation memory

Pass the threadId returned by the first run into follow-up runs to keep conversation context:

{ "query": "Now compare it with the second-best target.", "threadId": "abc123..." }

Server configuration

Env var	Effect
PORT	listen port (default 3000)
AGENTPACK_RUN_LIMIT	runs per IP per hour (default 30, 0 = off)
AGENTPACK_BUILD_LIMIT / AGENTPACK_SUGGEST_LIMIT	builder rate limits (default 10/hour)
AGENTPACK_DYNAMIC=0	disable browser-built teams
AGENTPACK_MODEL / AGENTPACK_SUPERVISOR_MODEL	model overrides